Abstract:
Data have been growing enormously in various domains including society, eco nomics, industry, security, transportation, and medicine. The high dimensional struc ture of these data requires new techniques that employ their underlying connectivity structure. Graph signal processing (GSP) has emerged as a processing tool for high dimensional datasets as an extension of classical signal processing performed in the Euclidean space. In this thesis, electroencephalography (EEG) data collected for brain computer interfacing (BCI) are used for classification using GSP as a preprocessing tool. Two EEG datasets, one during emotion detection, and one during motor imagery are used. Support vector machines (SVM) and K−nearest neighboring algorithms are used for classification. The underlying connectivity structure of the EEG data is ob tained using the distance and neighboring information of the electrode locations on the scalp. The results show that the classification accuracy is significantly improved when the data are projected to the underlying graph subspace determined by the graph spectral eigenvectors followed by a temporal filtering determined by Fourier spectral eigenvectors as a preprocessing step before classification.|Keywords : Graph Signal Processing, EEG, SVM, KNN, Brain Computer Interfacing (BCI).